Instruction–Level Parallelism VLIW, Vector, Array and Multithreaded

Chapter 07: Instruction–Level Parallelism─ VLIW, Vector, Array and Multithreaded Processors …

Lesson 02: Pipeline level and higher level Parallelism Concepts in Parallel Processing Objective

• To learn effects of parallel pipelines • To learn superscalar processors • Understand pipeline and higher levels of parallelisms

• Execute several instructions is parallel • Two or more instructions execute in parallel as well as in pipeline • The instructions that make up a program are handled in superscalar processors by the instruction issue logic

• The logic issues instructions to the units in parallel • This allows control flow changes, such as branches, to occur simultaneously across all of the units, making it much easier to write and compile programs for instruction-level parallel superscalar processors

• Performance further improves in processors by allowing independent instructions to execute simultaneously (called instruction-level parallelism)

Schaum’s Outline of Theory and Problems of Computer Architecture 8 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 Pipelining and instruction-level parallelism combine in the superscalars

• Most modern processors (called superscalars) employ both techniques to improve performance

• SunSparc • Pentium, • PowerPC •ARM

Schaum’s Outline of Theory and Problems of Computer Architecture 10 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 Parallel processing techniques to exploit instruction-level parallelism

• Superscalar • VLIW processing • Vector processing • Array processing

Schaum’s Outline of Theory and Problems of Computer Architecture 11 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 Instruction Issue Logic and Four execution units in a superscalar

• VLIW instruction-level parallelism relies on the compiler to determine which instructions may be executed in parallel and providing that information to the processor hardware (instruction issue logic)

• Require that programs be recompiled for the new architecture, but achieve very good performance on programs written in sequential languages such as C or FORTRAN when these programs are recompiled for a VLIW processor

• The operations such as multiply are first divided into several steps and a stream of operands (called vectors) operated in for each step parallel processing units

• A vector processor functions for instruction level parallelism by using instruction-parallel architecture along with vector element operands in the parallel processing pipelines • It is most common special case of pipelining

• The operations are first divided into several streams of operands (called array elements) and run for each stream in the parallel processing units • It is a special case of pipelining, where the parallel pipelines operate on different streams of operands • An array processor functions as data parallelism by using data-parallel architecture. There is single instruction processing multiple data in parallel processing units

• An array processor functions as data parallelism by using data-parallel architecture • Single instruction processing multiple data in parallel processing units

• The operations─ first divided into several threads, and then each thread is divided into instruction streams • Instruction streams in parallel • Concurrent execution of thread • When a thread is blocked or delayed then another thread instruction-streams run in the pipelines

• Can run different threads in time slicing mode or in priority mode • A special case of parallel processing, where the parallel pipelines operate on instruction streams of a thread

Schaum’s Outline of Theory and Problems of Computer Architecture 20 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 Multiprocessor Systems─ Highest level Coarse-grained Parallelism

• Multiprocessors give parallelism at processor level • An application divided into number of processes • Each process may be broken into threads each of which processes by two or more processors or processor cores on single chip

Schaum’s Outline of Theory and Problems of Computer Architecture 21 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 Multiprocessor Systems─ Highest level Coarse-grained Parallelism

• Processors may be identical or heterogeneous and distributed • Inter-process communication between different processors • The ration of time spent in computation on one processor << the inter-process communication time

• Parallel Pipelines improves Performance • Superscalars has special hardware techniques to enable issue of instructions after resolving conflicts by scheduling the instructions dynamically and by predicting branches by hardware in superscalar • VLIW processor • Vector processor • Array processor

• Multithreaded processor • Multiprocessor system

Schaum’s Outline of Theory and Problems of Computer Architecture 25 Copyright © The McGraw-Hill Companies Inc. Indian Special Edition 2009 End of Lesson 02 on Pipeline level and higher level Parallelism Concepts in Parallel Processing